2012 International Conference on Electronics Engineering and Informatics (ICEEI 2012)
IPCSIT vol. 49 (2012) © (2012) IACSIT Press, Singapore
DOI: 10.7763/IPCSIT.2012.V49.32
An Advanced Detection Scheme for Concealed Micro-Electronic
Devices Using Classification Algorithms
Kwangyul Kim1, Jeonghwan Lim2, Songkang Kim2, Junkyung Cho2, and Yoan Shin1
1
School of Electronic Engineering, Soongsil University, Seoul 156-743, Korea
2
R&D Center, Eltronix Co., Seoul 152-769, Korea
Abstract. This paper proposes an advanced detection scheme for concealed electronic devices using
classification algorithms in order to protect intellectual property rights. In general, if a radio wave with a
specific fundamental frequency is propagated from the transmitter of a classifier to a concealed object, the
second and the third harmonics will be returned as the radio wave is reflected. Using this principle, we use
the k -NN (Nearest Neighbor) and the fuzzy c-means algorithm for judging diverse concealed objects. The
simulation results indicate that the proposed scheme can detect electronic devices and metal devices in
diverse learning data environments through efficient classification. Furthermore, it is expected that the
proposed schemes can be extended to accumulate learning data of diverse devices in order to use the
proposed scheme for more efficient detection of diverse concealed objects to protect intellectual property
rights.
Keywords: concealed micro-electronic devices, harmonics, detection, classification, decision boundary
1. Introduction
Recently, as the knowledge-based IT society has emerged due to the innovative development of the
internet and wireless mobile communications, the importance of intellectual property rights including patent
rights, has been gradually magnified as competition among various research and development has become
fierce. For instance in Korea, the number of applications for intellectual property rights has been rapidly
increasing every year since 1986. Based on these intellectual property rights, technical competition has
become fierce not only among businesses in Korea but also throughout the world. As a result, according to
a report by the World Bank published in 2009, the GDP (Gross Domestic Product) of Korea was ranked
fourteenth in the world, and an IMD (Institute for Management Development) report indicated that Korean
scientific technology took third place in the world, while Korean technical competitiveness took fourteenth
place in the world [1].
However, although the number of intellectual property rights increased as the importance of intellectual
property rights as such has been magnified, the outcomes of technology development have not been
efficiently managed and protected. In particular, the 2009 IMD report indicated the level of cyber security of
Korea as thirty-eighth in the world, and the 2009 WEF (World Economic Forum) reported the level of
protection of intellectual property rights of Korea was forty-first in the world [1]. Thus, whereas the level of
research and development has been enhanced, the contents of research and development should be
effectively protected; as yet, efforts to this end are quite insufficient. As a result, industrial secrets have
frequently been leaked, and economic damage has been increasing. In particular, since large data can be
stored recently in micro storage media such as memory sticks, the possibility that insiders may leak
documents more easily by concealing micro devices has increased.

Yoan Shin Tel: +82-2-820-0632; Fax: +82-2-821-7653;
E-mail: yashin@ssu.ac.kr)
175
As a method to tackle this issue, a proper scheme is proposed in [2] that can judge the presence of
electronic devices based on the characteristics of the second and third harmonics returned from concealed
objects when the transmitter has transmitted a radio wave of a specific fundamental frequency. That scheme
will judge the object to be an electronic device if the magnitude of the second harmonics is larger than that of
third harmonics and to be metals in the opposite cases. However, due to the effects of interference by internal
and external electromagnetic waves that may occur in the process of measuring concealed objects, false
alarms may occur if the objects are judged using only magnitudes of harmonics.
Hence, in order to efficiently reduce detection errors resulting from interferences by internal and external
electromagnetic waves, k -NN (Nearest Neighbor) and fuzzy c-means algorithms will be used in this paper.
Through computer simulations, it could be observed that the present scheme was more efficient than general
schemes since it showed detection error rates of around 10%, even in electromagnetic wave interference
environments. The rest of the paper is organized as follows. In Section 2, the classification algorithms
considered in the proposed detection schemes are presented. The proposed schemes are introduced in Section
3, followed by the simulation results in Section 4. Finally, Section 5 concludes the paper.
2. Classification Algorithms
2.1. k -NN Algorithm
A classification algorithm used in this paper is a k -NN algorithm which is a classical data classification
algorithm to assume the classes of new data based on learning data of which the classes have already been
defined [3]. In particular, this algorithm is widely used in diverse machine learning areas since its usefulness
and conciseness have been proved. In the case of general k -NN classification methods, Euclidean distances
between predefined learning data and new data to be classified are calculated as in (1) to find k pieces of
nearest learning data, and the measured new data are defined into the class that contains more learning data
among the k pieces of learning data.
D  ( x1  y1 ) 2  ( x2  y2 ) 2    ( xd  yd ) 2
(1)
where xi  i  1,, d  is the attribute values (i.e., features) of the pre-defined learning data and
yi  i  1,, d  is the attribute values of the object of which the class has not been defined.
If we have a sufficient number of samples, the majority vote of the nearest k neighbors may be better
than a single nearest neighbor. The value of k should be large enough to minimize the probability of a
non-Bayes decision for the unclassified point x, but at the same time it should be small enough in order that
the points be close enough to x to give an accurate estimate of the posterior probabilities of the true class of
x [3].
2.2. Fuzzy c-means Algorithm
Fuzzy c-means algorithm is a data-clustering technique in which a data set is grouped into n clusters
with every data point in the data set belonging to every cluster to a certain degree. For example, a certain
data point that lies close to the center of a cluster will have a high degree of belonging or membership to that
cluster, while another data point that lies far away from the center of a cluster will have a low degree of
belonging or membership to that cluster. This is based on minimization of the following objective function as
in (2) [4,5].
N
C
J m   uijm xi  c j , 1  m  
2
(2)
i 1 j 1
where m is any real number greater than 1, a d -dimensional vector x i is the i -th  i  1,, N  measured
data, c j is the d -dimensional center of the cluster, uij is the degree of membership of x i in the cluster j
 j  1,, C , and
 is any norm expressing the similarity between the measured data and the center. Fuzzy
176
partitioning is carried out through an iterative optimization of the objective function shown above, with the
update of membership uij and the cluster centers c j by (3).

 C  xi  c j
uij    

 k 1  x i  c k





2
m 1
1


 ,


N
cj 
u
x
m
i j
i 1
i
N
u
i 1
(3)
m
ij
This iteration will stop when max uijk 1  uijk    , where  is a termination criterion between 0 and 1,
ij
whereas k is the iteration steps. This procedure converges to a local minimum or a saddle point of J m .
3. Proposed Detection Scheme
3.1. Detection of Micro-Electronic Devices
When the transmitter of a classifier has transmitted a radio wave of a specific fundamental frequency, the
signals returned from concealed objects contain harmonics of which the frequency is n times the
fundamental frequency when they are received. These harmonics have diverse characteristic values,
depending on target objects. In general, in the case of electronic devices, the peak intensity of the second
harmonics is higher than that of the third harmonics; conversely in the case of metal devices, the peak
intensity of the third harmonics is higher than the peak intensity of the second harmonics [2]. Hence, in this
paper, electronic devices and metal devices are distinguished by comparing the second and the third
harmonics returned and received. The present section offers explanation on the environments that occur
when learning data are measured in order to classify products.
Fig. 1: Asymmetric learning data and D/B.
Figure 1 shows a situation where the learning data of the second and the third harmonics are
asymmetrical. It can be intuitively seen that, if a symmetrical D/B (Decision Boundary) as shown by the thin
dotted line is used when learning data are asymmetrically located, most devices may be determined to be
electronic devices, and thus CEP (Classification Error Probability) may increase. In addition, in the cases
where learning data are symmetrical or asymmetrical, an appropriate D/B should be set so that optimum
classification can be made based on the D/B to efficiently reduce CEP resulting from interfering
electromagnetic waves. Thus, considering the average values of the learning data of electronic devices and
metal devices, an advanced D/B as shown by the solid line in the figure has been proposed and set. In
particular, it is expected that this will enable more precise classification, even in environments with noises
that may occur in the process of measuring learning data or interferences by internal and external
electromagnetic waves that may occur in the process of measuring new data.
177
However, to provide optimum D/B’s for environments with asymmetrical learning data, symmetrical
D/Bs should be set based on learning data. Therefore, in this paper, a D/B scheme using fuzzy c-means
algorithm was proposed to enable efficient classification, even in environments where learning data are
asymmetrical.
3.2. Proposed D/B and Detection Scheme
The proposed detection scheme is composed of the following two steps. The first step is to obtain
learning data and define classes based on defined D/B’s using the fuzzy c-means algorithm, and the second
step is to classify classes for new devices using the k -NN algorithm.
3.2.1.
First Step: Fuzzy c-means Algorithm Based Decision Boundary Setting
The purpose of the first step is to make virtual symmetrical D/B’s for accurate classification in
environments where learning data are asymmetrical. Thus, the learning data are classified using a fuzzy
c-means algorithm first. In this case, virtual learning data are used. The virtual learning data are distributed
on average points of learning data values in two classes, and the fuzzy c-means algorithm classify the classes
using the learning data and the virtual learning data. Hence, as shown in Fig. 1, a symmetrical D/B for the
asymmetrical learning data will be formed in the center of the space between the two classes. In Fig. 1, the
thin dotted line is a “Conventional D/B”, the solid line is an “Advanced D/B”, and finally the thick dotted
line indicates the “Proposed D/B”. That is, it can be seen that the “Proposed D/B” is a symmetrical D/B set
considering the asymmetrical learning data.
3.2.2.
Second Step: Classification and Detection
The second step is to classify the new data and check if the data indicate electronic devices. For the
classification, the k -NN algorithm explained in Section 2 is used. By calculating the Euclidean distances as
shown in (1) between newly measured data and individual learning data, k pieces of learning data nearest to
the measurement data are found. Thereafter, the measured data are classified into the class that contains the
largest number of learning data. In this case, if the class designated in advance in the first step is the same as
the class determined in the second step, the detection will be considered to have been properly conducted; if
not, the detection will be considered as involving errors.
4. Simulation Results
To measure the performance of the proposed detection scheme for the electronic devices, computer
simulations were conducted using MATLAB. The number of learning data used was 100, consisting of 50
electronic devices and 50 metal devices; the classes of the electronic devices and metal devices were defined
using all three types of D/B’s explained earlier. Since the scheme is intended to reduce CEP resulting from
interferences by internal and external electromagnetic waves, learning data into which interferences were
intentionally included, were used. At this time, the interferences into the learning data had the characteristics
of a Gaussian distribution with the mean 0 and the standard deviation of 50. The measurement data were
created to have the characteristics of a Gaussian distribution against the D/B that determined two classes.
The standard deviation of the created measurement data was set to 30. Since the measurement data could
also be interfered by internal and external electromagnetic waves in the process of measurement, a Gaussian
random variable (with standard deviation of 2) was injected into the original measurement data. The k
values of the k -NN algorithm were increased by 4 pieces at a time in a range of 3~51 during the
measurement; the measurement was conducted 100,000 times.
Figure 2 shows CEPs according to k in environments where learning data are asymmetrical. It can be
observed that, in the case of the “Conventional D/B”, CEP’s of around 50% were obtained since efficient
classification is possible only in symmetrical environments. However, in the cases of all schemes other than
the “Conventional D/B”, CEP’s decreased as k increased. CEP’s decreased by around 2% in the case of the
“Advanced D/B” and around 5% in the case of the “Proposed D/B”. When the schemes were compared for
k = 3, CEP’s decreased further by around 6% in the case of the “Proposed D/B” than in the case of the
“Advanced D/B”; in cases where k = 51, CEP’s decreased further by around 8%. In particular, in the case
of the “Proposed D/B”, the excellence of the proposed scheme could be identified. The reason is that
178
environments where learning data were asymmetrical were changed into virtual symmetrical environments to
obtain results similar to the results from the k -NN algorithm, and that relative margins were given to spaces
between the learning data to enable accurate classification, even in environments with the interferences by
internal and external electromagnetic waves.
50
Conventional D/L
Advanced D/L
Proposed D/L
45
Error Probability (%)
40
35
30
25
20
15
10
5
0
3
7
11
15
19
23 27 31
Number of k
35
39
43
47
51
Fig. 2: CEP performance comparison of the proposed D/B with the conventional D/B and advanced D/B in the
asymmetric environments.
5. Conclusion
This paper proposed an advanced detection scheme to improve the classification performance of micro
electronic device detectors being developed in order to protect the intellectual property rights. To measure
the performance of the proposed scheme, diverse D/B’s that could define learning data classes in advance
were set; and based on these D/B’s concealed objects were classified using the k -NN algorithm. In
particular, cases where the second and the third harmonics of learning data were considered and the effects
of interferences by internal and external electromagnetic waves that might occur in the process of detecting
concealed objects were also considered. From the computer simulations, it was observed that the proposed
D/B scheme provided CEP reductions of around 9% in asymmetrical environment. Hence, it could be seen
that the proposed scheme could not only detect concealed microelectronic devices but also show efficient
classification performance, even under the effects of interference by internal and external electromagnetic
waves.
6. Acknowledgements
This work was supported by the Industrial Convergence Source Technology Development Program (No.
10039192) funded by the MKE, Korea.
7. References
[1] Korea Industrial Technology Association. A Study on the Current Status of Overseas Industrial Security
Management and Case Studies of American Companies. Small and Medium Business Administration of Republic
of Korea, 2009.
[2] B. R. Barsumian, T. H. Jones, and C. A. Liter. Surveillance device detection utilizing nonlinear junction detection
and reflectometry. US Patent #7,212,008, 2007.
[3] T. M. Cover, and P. E. Hart. Nearest neighbor pattern classification. IEEE Trans. Inform. Theory. 1967. IT-13 (1):
21-27.
[4] F. Palumbo, C. N. Lauro, and M. J. Greenacre. Data Analysis and Classification. Springer, 2010.
[5] S. Theodoridis, and K. Koutroumbas. An Introduction to Pattern Recognition: A MATLAB Approach. Academic
Press, 2010.
179